cdc-error-plugin

Design, review, or refactor CLI error messages so they serve BOTH human operators AND LLM agents using the RFC 9457 Problem Details envelope with mandatory agent extensions (retry_after, suggested_fix, code_actions[]). Use whenever the user works on CLI error output, stderr format, exit codes, machine-readable errors, structured error envelopes, dual-format CLI rendering, or reducing token cost of tool_result payloads. Triggers on phrases like "CLI error message", "stderr format", "machine-readable errors", "dual-consumer errors", "problem+json", "RFC 9457 errors", "agent-parseable errors", "review my CLI error messages", "refactor error handling", or when improving error reporting in any CLI (Go, Rust, Python, Node, Ruby, Java, C#, shell). Use this skill even when the user does not cite the RFC or dual-consumer framing — if the work touches CLI error output an agent might parse, this skill applies.

Review, refactor, or design error-handling source code across any language (Rust, Go, Python, TypeScript, Java, others) and any context (CLI, library, service, web). Produces classified violation reports (must-fix / should-fix / nit / question / praise), before/after diffs, and YAML work plans for large refactors. Triggers on "review my error handling", "audit error code", "find error-handling bugs", "fix exception handling", "improve error propagation", "review my Result usage", "audit my errors.Is calls", "check my try/except", "panic in library", "swallowed error", "lost cause chain", or when reviewing code where modified files contain error emission, propagation, recovery, or cleanup logic. Defers CLI envelope and format questions to sibling skill cdc-err (RFC 9457 Problem Details). Use even when the user does not request a review — if the work touches how errors propagate, get swallowed, get wrapped, or get logged, this skill applies.

Interpret and act on RFC 9457 problem+json error envelopes that a CLI (typically one shaped by sibling skill cdc-err) returns to an LLM agent in a tool_result. Three modes — Parse the envelope, Decide retry vs apply-fix vs handback vs abort, Act by emitting a concrete executable action plan. Always gates suggested_fix and code_actions[] on applicability markers (machine_applicable / maybe_incorrect / has_placeholders / unspecified) and treats missing/unspecified applicability as maybe_incorrect (safe default — surface to human, never auto-apply). Use this skill whenever a tool_result payload contains an application/problem+json envelope, fields like retry_after / suggested_fix / code_actions[], or status / title / type / detail / instance combinations characteristic of RFC 9457. Also use on explicit phrasing — "interpret this error", "should I retry this CLI error", "the CLI returned problem+json", "what does this retry_after mean", "can I auto-apply this suggested_fix", "the tool_result has an error envelope